Implement NL2Code training for Hearthstone dataset

[rshin]

• [x] Finish token generation mechanism
• [x] Create encoder
  • [x] Copying tokens
  • [x] Attention
• [x] Create code to read data
• [x] Bahdanau attention
• [x] Generate vocabularies for source and target Depending on model, we need different vocabularies Look at OpenNMT Look at tensor2tensor
• [x] Blacklist some elements in Python grammar ("ctx" fields)
• [x] Add optimizers to registry
• [x] Improve registry to avoid config
• [x] Setup model to connect everything together (enc2dec)
• [x] Figure out how to specify vocabulary, grammar, etc. to models
• [x] Figure out what to do when a singular type has only one derivation: Nothing special in the case of no unary closure
• [x] Masking of actions
  • [x] At training time: no masking applies
• [x] Deal with types of constants: Num -> object should be Num -> int, Num -> float
• [x] Name -> identifier -> str to Name -> str
• [x] Variational dropout Complicated to do in PyTorch
• [x] Ensure that all loss components have positive sign
• [x] Adjust grammar based on empirical observations throughout the data
• [ ] Implement unary closures

Others

• [ ] Save training progress more permanently
• [ ] Batching + constructing training instructions in separate processes
• [ ] Introduce abstract base classes
  • [x] preproc
  • [ ] model
  • [ ] encoder